# Skywalkers - **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon) - **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST) - **Where:** Online - **Team:** [Abhay Tiwari](https://cerebralvalley.ai/u/tabhay) - **GitHub:** https://github.com/Skywalkers-24/LoadTruth - **Demo video:** https://youtu.be/ilH6tjTAGKs - **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery - **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/456 LoadTruth: Every Load. Every Measurement. Every Dollar. Independently Verifiable. Every disaster debris load ticket represents public money that should be independently auditable. Yet today, much of that money is settled using subjective visual estimates that cannot be reliably reproduced after the debris is gone. Following major hurricanes, municipalities award multi-hundred-million-dollar emergency debris removal contracts. Haulers are typically compensated by the cubic yard (CY) of debris delivered to temporary staging sites. Field monitors estimate truck-bed fill percentages from elevated platforms, recording subjective measurements on paper or handheld applications. This creates a significant verification gap. Historical DHS Office of Inspector General audits have documented weaknesses in debris monitoring, including unsupported quantities, duplicate billing, and invoices exceeding certified vehicle capacities. The consequences include disputed payments, millions in questioned costs, and FEMA reimbursement clawbacks years after the original operations. LoadTruth closes that verification gap by turning ordinary inspection cameras and field radio recordings into a reproducible, evidence-backed settlement system. How LoadTruth Works 1. Spatial Grounding with Meta SAM 3.1 LoadTruth uses Meta SAM 3.1 to segment truck-bed boundaries, tailgate reference planes, and debris surfaces across video frames. It generates geometric masks and load contours, accounting for perspective and camera positioning through calibrated reference geometry. Multiple valid frames are aggregated to reduce the impact of occlusion, motion, and inconsistent visibility. 2. Multimodal Intelligence with Meta Muse Spark Meta Muse Spark analyzes visual and contextual evidence to extract vehicle identification and certified capacity information from truck placards, classify debris into FEMA-relevant categories (Vegetative, Construction & Demolition, White Goods, and Hazardous), and identify suspicious operational patterns. This includes potential duplicate trips, implausibly short turnaround times, capacity violations, and visual indicators of hollow or partially filled loads. 3. Acoustic Verification with Meta Muse Voice LoadTruth processes noisy VHF field-radio recordings, transcribing monitor callouts and matching reported fill estimates against video-derived measurements. Discrepancies between verbal declarations and visual evidence are flagged for human review. 4. Deterministic Financial Settlement with No LLM-Generated Billing AI models never determine the final payable volume or authorize payments. LoadTruth combines validated visual measurements with certified vehicle dimensions and capacity records. A deterministic Python calculation engine uses exact Decimal arithmetic to calculate billable quantities, apply configured validation rules, and flag discrepancies before settlement. Every calculation is traceable to its underlying measurements and source evidence. 5. Cryptographic Evidence for One-Click Independent Auditing Every load generates an evidence package containing references to the original video, selected frames, segmentation masks, model outputs, measurement data, and settlement calculations. SHA-256 fingerprints detect changes to the recorded evidence. Auditors can retrieve the original inputs, independently verify their integrity, and reproduce the settlement calculation without rerunning AI inference. 6. Resilient Field Operations Built for Disaster Zones Disaster operations cannot depend on perfect connectivity. LoadTruth is designed to support local GPU inference and evidence capture during network outages, with synchronization when connectivity returns. The Outcome LoadTruth transforms debris monitoring from a subjective, paper-based process into a measurable and independently verifiable financial control. • For municipalities: Greater visibility into contractor billing and public expenditure. • For field monitors: Automated measurements and discrepancy detection instead of manual estimation alone. • For auditors: Reproducible evidence trails without recurring AI inference costs. • For FEMA reimbursement: Stronger supporting documentation for quantities submitted for reimbursement. The fundamental principle is simple: AI measures. Deterministic systems calculate. Humans review. Cryptographic evidence preserves the truth. LoadTruth doesn't ask auditors to trust an AI-generated number. It gives them the evidence to verify it themselves. ## More from Meta Global AI Developer Hackathon - [Malchy](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/453) - [Cookie-Team](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/454) - [NoMouse](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/455) - [Dedudev](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/457) - [BeKavod](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/458) - [Doc'O Clock](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/459) --- Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/456. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt